PAI official images

Overview

PAI provides two types of Python images:

1. CPU images: Built on the official Ubuntu base image for CPU computing.

2. GPU images: Built on the official CUDA base image for GPU computing.

Key features

1. Supports Ubuntu 22.04 and Ubuntu 24.04.

2. Supports Alibaba Cloud’s high-performance RDMA networking.

3. Supports Python versions 3.10 through 3.14.

4. Supports CUDA versions 12.4 through 13.0.

5. Includes common development tools such as curl, git, wget, rclone, and ping.

6. Uses Alibaba Cloud mirrors for pip and apt.

Overview

PAI provides two types of PyTorch images:

1. Built on PAI Python images with PyTorch, TorchVision, and TorchAudio packages pre-installed. These images inherit all the features of the Python images and cover official PyTorch releases from version 2.4.0 onward.

2. Built on NVIDIA NGC PyTorch images, with common development tools pre-installed and Alibaba Cloud mirrors configured for pip and apt.

Tag descriptions

-accl:

1. These images are pre-installed with the Alibaba Cloud High-Performance Collective Communication Library (ACCL). ACCL delivers higher communication performance than NCCL.

2. When developing or training with ACCL-based images, you should use the preconfigured Python environment. If you want to use a Python virtual environment, follow the installation guide to configure ACCL in your environment.

-ngc:

1. These images are built from NVIDIA NGC PyTorch images. The tag includes the NGC version. For example, 2.10.0-gpu-py312-cu130-ubuntu24.04-ngc25.11 is based on NGC PyTorch 25.11.

2. NGC PyTorch image features are documented in the NVIDIA official documentation.

Overview

Data-Juicer is a Ray-based distributed framework for data cleaning and preprocessing that enhances data quality for LLM training. PAI provides two image types:

1. CPU images: Built on PAI’s CPU base image for large-scale CPU-only tasks such as text processing and data cleaning.

2. GPU images: Built on PAI’s CUDA base image for GPU-accelerated tasks such as model inference and quality scoring.

Key features

• Based on Ubuntu 22.04.

• Supports Alibaba Cloud RDMA for high-throughput, low-latency distributed data loading and processing.

• Includes a full Data-Juicer runtime with built-in processors and Ray Dashboard monitoring.

• Supports CPU/GPU heterogeneous resource scheduling for diverse workloads such as data cleaning, quality evaluation, and multimodal data generation.

• Uses Alibaba Cloud mirrors for pip and apt by default.

Overview

Responsible AI ensures safety, reliability, fairness, transparency, and compliance throughout the AI model lifecycle. PAI provides two base images for Responsible AI:

1. CPU images: Built on the official Ubuntu image for general CPU computing, with Responsible AI toolchains integrated.

2. GPU images: Built on the official CUDA image for high-performance GPU scenarios, with Responsible AI toolchains integrated.

Key features

1. Supports Ubuntu 22.04 images.

2. Supports Python versions 3.11 through 3.14.

3. Supports CUDA 11.8.

4. Includes Responsible AI visual analytics tools with an interactive dashboard for multidimensional analysis, such as model fairness and error analysis, to help developers identify potential bias and errors.

5. Supports differential privacy training to prevent sensitive data leakage and meet compliance and privacy requirements.

6. Includes the RAI model encryption SDK (RAI_SAM_SDK) for sharded encrypted storage of LLMs and authorized decryption during inference.

Overview

Ray is a high-performance distributed computing framework for ML training, hyperparameter tuning, reinforcement learning, and inference. PAI provides two image types with ray[default] pre-installed (includes Ray Dashboard and common runtime components):

1. CPU images: Built on PAI’s CPU base image for CPU-only distributed computing and data processing.

2. GPU images: Built on PAI’s CUDA base image for GPU-accelerated training, inference, and large-scale parallel computing.

Key features

• Based on Ubuntu 22.04 and Ubuntu 24.04.

• Supports Alibaba Cloud RDMA for high-throughput, low-latency distributed communication.

• Includes a full Ray runtime environment with common components that allow you to quickly start Ray Head and Worker nodes and run tasks.

• Supports CPU/GPU heterogeneous resource scheduling for diverse workloads such as training, data processing, and inference.

• Uses Alibaba Cloud mirrors for pip and apt by default.

Overview

The ModelScope Library handles model and dataset management, training, and inference with PyTorch, TensorFlow, and other frameworks (Python 3.8+, PyTorch 1.11+, TensorFlow). ModelScope official images let you skip environment setup.

Overview

TorchEasyRec is a deep learning framework for recommendation systems covering matching, ranking, multitask learning, and generative recommendations.

PAI provides official TorchEasyRec images pre-installed with dependencies such as pytorch, torchrec, fbgemm, and tensort. Two image types are available:

1. GPU version: Built on Ubuntu 22.04 with CUDA acceleration for high-performance large-scale recommendation model training (recommended).

2. CPU version: Built on Ubuntu 22.04 for development, debugging, and small-scale training (Note: Some operations are GPU-only).

Overview

Triton Inference Server is a high-performance NVIDIA inference server supporting TensorFlow, PyTorch, ONNX Runtime, and other frameworks through a consistent serving interface.

Key features

1. Multi-framework support: Triton Server supports various deep learning frameworks and model formats, enabling unified deployment of diverse models.

2. High throughput and low latency: Triton improves inference performance through batching and parallel inference. It also leverages NVIDIA GPU acceleration to maximize compute power.

3. Dynamic model management: Triton allows dynamic loading and unloading of models, which enables flexible version control, A/B testing, and model updates.

4. Simple APIs and scalability: Triton offers REST and gRPC interfaces for easy integration. It also integrates seamlessly with container orchestration systems such as Kubernetes for large-scale inference deployments.

5. Heterogeneous hardware support: In addition to NVIDIA GPUs, Triton runs on CPUs and other accelerators, which supports deployment across diverse hardware platforms.

6. Custom post-processing: Users can apply custom logic to inference results to meet specific application needs.

Overview

ComfyUI is a node-based GUI for running and customizing diffusion models such as Stable Diffusion. Users drag and drop components to build image generation pipelines without code.

Key features

1. Node-based workflow: Breaks down steps such as text encoding, sampling, model loading, and image post-processing into independent nodes that users can freely connect for precise control.

2. Efficient resource management: Loads only the models needed for the current workflow, which reduces VRAM usage and supports batch generation and complex pipeline optimization.

3. Highly extensible: Supports custom node plugins with a rich community ecosystem, such as ControlNet, LoRA, and Upscale, for easy integration of new models or features.

4. Workflow export and sharing: Entire generation workflows can be exported as JSON files for reproducibility, collaboration, or deployment to other environments.

Overview

PAI-RAG is an enterprise-grade RAG conversational system built on PAI-EAS. It integrates LLMs with knowledge retrieval for private knowledge Q&A and intelligent customer service. An open-source modular framework (GitHub: aigc-apps/PAI-RAG) is available for customization.

Key features

• Multiple vector database support: Natively compatible with Elasticsearch, Hologres, Tablestore, Milvus, and other mainstream vector databases to meet diverse enterprise needs.

• Web search enhancement: Supports real-time web retrieval to overcome the timeliness limitations of model pretraining data and improve answer accuracy and freshness.

• Flexible deployment and integration: Offers a WebUI, RESTful API, and OpenAI-compatible interface for quick integration into existing business systems.

• End-to-end knowledge base management: Supports document upload and management using the WebUI or OSS, with one-stop capabilities for chunking, vectorization, version updates, and knowledge base operations.

Overview

vLLM is an open-source LLM inference and serving engine. It uses advanced memory management and scheduling to deliver high throughput with low latency.

Key features

1. PagedAttention: A core innovation inspired by OS paging mechanisms to dynamically manage KV Cache, eliminate VRAM fragmentation, and significantly improve VRAM utilization efficiency.

2. Continuous batching: Dynamically merges requests of varying lengths for parallel decoding, which greatly improves GPU utilization and throughput.

3. High throughput, low latency: Supports higher concurrency on the same hardware and is ideal for high-traffic production environments.

4. Developer-friendly: Provides a simple Python API and an OpenAI-compatible interface for rapid integration into existing applications.

5. Rich ecosystem: Natively supports advanced features such as LoRA fine-tuning inference, multimodal models, and tool calling (Function Calling).

Overview

EasyAnimate is a PAI-developed video generation framework based on Diffusion Transformer (DiT). It generates high-quality videos from text or images and covers data preprocessing, VAE training, and DiT inference.

Key features

• High-resolution long video generation: Generates coherent videos up to 1024×1024 resolution and 6 seconds or longer.

• Multimodal input: Supports both text prompts (text-to-video) and image inputs (image-to-video) for dynamic video generation.

• Complete training pipeline: Offers end-to-end training capabilities for VAE, DiT base models, and LoRA fine-tuning to support customized development.

• Production-ready deployment: Officially supported by PAI inference services for seamless integration into cloud inference platforms and is suitable for production environments.

Overview

Kohya is a Gradio-based toolset for Stable Diffusion fine-tuning with LoRA, DreamBooth, and other techniques.

Key features

• Multiple training methods: Natively supports LoRA, DreamBooth, full-parameter fine-tuning, and SDXL model training.

• Graphical interface: Provides an intuitive Web UI (based on Gradio) where users can configure parameters using forms instead of command-line coding.

• Cross-platform compatibility: Primarily designed for Windows but also supports Linux and macOS.

• End-to-end toolchain: Integrates data preprocessing, auto-captioning, training monitoring, and model export to cover the full fine-tuning lifecycle.

• Open source and active community: Fully open-source with continuous community maintenance and compatibility with mainstream inference frameworks, such as Stable Diffusion WebUI, for direct deployment of trained models.

Overview

Stable-Diffusion-WebUI is an open-source GUI for deploying and running Stable Diffusion models locally, supporting text-to-image and image-to-image generation.

Key features

• Multimodal generation: Supports mainstream modes such as text-to-image (txt2img), image-to-image (img2img), inpainting, and outpainting.

• Rich extension ecosystem: The built-in plugin system supports popular extensions such as ControlNet, LoRA, and T2I-Adapter to enhance generation control.

• Integrated training and fine-tuning: Includes DreamBooth, LoRA, and Textual Inversion for custom model fine-tuning.

• Cross-platform deployment: Runs on Windows, Linux, macOS, and Google Colab, and supports both CPU and GPU (NVIDIA/AMD) hardware.

• User-friendly: The web interface built with Gradio offers visual parameter configuration and is suitable for users from beginners to professionals.

Overview

CosyVoice is a high-fidelity speech synthesis model that clones a target voice from a prompt audio clip in under 30 seconds and supports cross-lingual voice replication. PAI-EAS packages it with an integrated WebUI for rapid deployment of cloud-based speech inference services.

Key features

• Zero-shot voice cloning: Replicates target voices from just 3 to 10 seconds of reference audio for personalized speech generation.

• Multilingual and cross-lingual synthesis: Supports Chinese, English, Japanese, Korean, and other languages while it maintains voice consistency across languages.

• Emotion and fine-grained control: Precisely controls vocal details such as emotion, laughter, and breathing through natural language descriptions.

• Highly human-like: Matches human speech in intonation, rhythm, and pauses, and significantly outperforms traditional TTS technologies.

• Real-time streaming synthesis: Supports low-latency streaming text-to-speech output for real-time interactive scenarios.

• Full-stack toolchain: Provides complete capabilities from inference and training to deployment for industrial-grade application integration.

Overview

SGLang (Structured Generation Language) is a high-performance LLM inference framework with a co-designed frontend language and backend runtime (SGLang Runtime) for controllable, low-latency, high-throughput model serving.

Key features

• Structured controllable generation: Natively supports precise output format control using JSON Schema, regular expressions, and other constraints to overcome the limitations of traditional prompt engineering.

• High-performance inference: Uses innovative optimizations such as RadixAttention and Radix Cache to achieve 3 to 5 times higher throughput than mainstream solutions.

• Multimodal support: Works with both text-only LLMs and vision-language models (VLMs), and supports multimodal inputs such as images and video.

• Flexible integration: Offers a simple Python API that can replace the OpenAI API for complex prompt workflows, which lowers development barriers.

Overview

TensorFlow Serving is a high-performance model serving system from the TFX ecosystem. It deploys TensorFlow SavedModel models as online inference services via gRPC and REST APIs.

Key features

• Model version management: Supports parallel loading of multiple model versions and seamless hot updates for phased releases and rollbacks.

• High-performance inference: Production-optimized scheduling and batching mechanisms ensure low-latency, high-throughput service.

• Out-of-the-box integration: Natively supports the TensorFlow SavedModel format without requiring additional conversion.

• Extensible architecture: Offers pluggable components such as Servable, Source, and Manager for custom loading logic and serving policies.

• Multi-protocol support: Provides both gRPC (high performance) and HTTP/REST (easy integration) interfaces to accommodate different client needs.